Print Close

FastSurfer-HypVINN: Automated sub-segmentation of the hypothalamus and adjacent structures on HiRes

Poster No:

1860

Submission Type:

Abstract Submission

Authors:

Santiago Estrada¹^,2, David Kügler¹, Emad Bahrami¹^,3, Peng Xu², Dilshad Mousa², Monique Breteler²^,4, N. Ahmad Aziz²^,5, Martin Reuter¹^,6^,7

Institutions:

¹AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE), Bonn, NRW, Germany, ²Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE), Bonn, NRW, Germany, ³Computer Science Deparment, University of Bonn, Bonn, NRW, Germany, ⁴IMBIE, University of Bonn, Bonn, NRW, Germany, ⁵Departmet of Neurology, University of Bonn, Bonn, NRW, Germany, ⁶A.A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, ⁷Department of Radiology, Harvard Medical School, Boston, MA

First Author:

Santiago Estrada
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)|Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE)
Bonn, NRW, Germany|Bonn, NRW, Germany

Co-Author(s):

David Kügler
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)
Bonn, NRW, Germany

Emad Bahrami
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)|Computer Science Deparment, University of Bonn
Bonn, NRW, Germany|Bonn, NRW, Germany

Peng Xu
Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE)
Bonn, NRW, Germany

Dilshad Mousa
Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE)
Bonn, NRW, Germany

Monique Breteler
Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE)|IMBIE, University of Bonn
Bonn, NRW, Germany|Bonn, NRW, Germany

N. Ahmad Aziz
Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE)|Departmet of Neurology, University of Bonn
Bonn, NRW, Germany|Bonn, NRW, Germany

Martin Reuter
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)|A.A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital|Department of Radiology, Harvard Medical School
Bonn, NRW, Germany|Boston, MA|Boston, MA

E-Poster

Introduction:

The hypothalamus is crucial in regulating a broad range of physiological, behavioral, and cognitive functions. Despite its essential role, only a few small-scale neuroimaging studies have investigated its substructures, primarily due to the lack of fully automated segmentation tools to address the scalability and reproducibility issues of manual segmentation. While a prior attempt to automatically sub-segment the hypothalamus with a neural network showed promise for 1.0 mm isotropic T1-weighted MRI [1], there is a need for an automated tool to sub-segment also high-resolutional (HiRes) MR scans (i.e. sub-millimeter), as they are becoming widely available. To this end, we recently proposed HypVINN [2], a fully automated deep learning method for the sub-segmentation of the hypothalamus and its adjacent structures on 0.8 mm isotropic T1-weighted (T1w) and T2-weighted (T2w) brain MR images that is robust to missing modalities (i.e., hetero-modal segmentation).

Methods:

HypVINN utilizes our novel hetero-modal Voxel-Size Independent Neural Network (HM-VINN) architecture at its core. The HM-VINN architecture allows the segmentation of multi-resolution scans in a single model. Furthermore, it also includes the capability of a hetero-modal model, permitting flexibility in the input modalities at inference time (i.e., segmentations can be generated using only a T1w or only a T2w or a T1w/T2w image pair). To train our tool, 44 manually segmented 0.8 mm isotropic scans from the Rhineland Study were used. Manual annotations were performed for eleven hypothalamic subregions and thirteen adjacent structures by an experienced rater using co-registered T1w and T2w images, as presented in Figure 1. HypVINN is extensively validated in terms of segmentation accuracy (dice similarity coefficient (Dice), volume similarity (VS), 95% Hausdorff distance (HD95)) against an unseen test-set, generalizability to 1.0 mm isotropic MRI scans from the UK Biobank (UKB), and in-session test-retest reliability (intra-class correlation (ICC), and VS). Furthermore, we tested the sensitivity of HypVINN to replicate known hypothalamic volume effects with respect to age and sex in two random subsets of the Rhineland Study (RS, n=463) at 0.8mm (HiRes) and the UK Biobank (UKB, n=535) at 1.0 mm.

Results:

HypVINN exhibits high segmentation performance both for standalone T1w images (Dice = 0.791, VS = 0.898, HD95 = 1.110 mm) as well as for T1w/T2w image pairs (Dice = 0.795, VS = 0.901, HD95= 1.086 mm). The proposed method can generalize remarkably well (Dice = 0.707, VS = 0.846, HD95= 1.718 mm) to 1.0 mm T1w scans from the UKB, an independent dataset never encountered during training with different acquisition parameters and demographics. Furthermore, HypVINN has an excellent test-retest agreement (ICC(A,1) > 0.95 and VS > 0.95) between in-session volume estimates. Finally, for the age and sex analysis, we observe that the global hypothalamic volume estimates and most sub-structures are negatively associated with age in both the RS and UKB subsets. Moreover, men have larger hypothalamic volumes than women, even after correction for head size (see Figure 2).

Conclusions:

HypVINN is the first hetero-modal deep learning method for hypothalamic sub-segmentation and the segmentation of other adjacent structures, such as the hypophysis, epiphysis, and major structures of the central optic system. In contrast to the only other contemporary automated method [1], our tool offers a more detailed parcellation of the hypothalamus. Furthermore, it can accurately generate segmentations from T1w and T2w MR images at isotropic resolutions of 0.8 mm or 1.0 mm. Finally, HypVINN will be integrated into the FastSurfer neuroimaging software suite, providing a user-friendly alternative for the reliable assessment of hypothalamic imaging-derived phenotypes in the neuroimaging community.

Modeling and Analysis Methods:

Methods Development ¹

Segmentation and Parcellation ²

Neuroanatomy, Physiology, Metabolism and Neurotransmission:

Subcortical Structures

Keywords:

Computational Neuroscience

Machine Learning

MRI

Segmentation

STRUCTURAL MRI

Other - Hypothalamus

^1|2Indicates the priority used for review

·Figure 1. T1-weighted (T1w) and T2-weighted (T2w) images and ground truth (GT) from two participants.

·Figure 2. Hypothalamic volume associations with age (a) and sex (b) in participants from the Rhineland Study (n=457) and UK Biobank (N=520) for HypVINN.

Provide references using author date format

1. Billot, B., Bocchetta, M., Todd, E., Dalca, A. V., Rohrer, J. D., & Iglesias, J. E. (2020). Automated segmentation of the hypothalamus and associated subunits in brain MRI. Neuroimage, 223, 117287.

2. Estrada, S., Kügler, D., Bahrami, E., Xu, P., Mousa, D., Breteler, M.M.B., Aziz, N.A., & Reuter, M. (2023). FastSurfer-HypVINN:
Automated sub-segmentation of the hypothalamus and adjacent structures on high-resolutional brain MRI. Imaging Neuroscience, Advance Publication. https://doi.org/10.1162/imag_a_00034

3.Henschel, L., Conjeti, S., Estrada, S., Diers, K., Fischl, B., & Reuter, M. (2020). Fastsurfer-a fast and accurate deep learning based neuroimaging pipeline. NeuroImage, 219, 117012.